High-Order Dependency Graph Learning for Container Anomaly Detection in Microservice Environments

Abstract

This paper addresses the structural modeling challenge of anomaly detection in modern containerized microservice systems. It proposes a graph attention network-based method for modeling container dependencies and detecting anomalies. The method builds a high-order dependency graph that represents dynamic invocation relationships between containers using runtime logs and monitoring metrics. It initializes node embeddings and employs a multi-head graph attention mechanism to adaptively model the interaction strength between upstream and downstream components. A multi-task optimization framework is introduced, including structural reconstruction loss, attention entropy regularization, and a contrastive objective. This enhances the model's expressiveness in terms of structural consistency and anomaly discriminability. The model can accurately aggregate key behavioral paths between containers and effectively model and score weak anomaly patterns within complex dependencies. To evaluate the method, multiple sensitivity experiments were designed to analyze performance trends concerning attention heads, embedding dimensions, and regularization parameters. The results confirm the adaptability and stability of the graph attention mechanism in container-level structural modeling and anomaly detection. The proposed method outperforms existing approaches across several key metrics and achieves accurate and robust anomaly identification in highly dynamic and complex system scenarios.